Video surveillance generates huge amounts of digital data continuously. It is difficult to monitoring manually. Video Anomaly Detection (VAD) address this challenge and by detect the events unsupervised manner. Due to the diversity of anomalous events, unsupervised learning methods have emerged as a practical solution. This paper presents a computationally efficient unsupervised video anomaly detection framework that leverages subsequent frame prediction using a lightweight ResNet-LSTM architecture. To optimize the robustness, a hybrid loss function is used which is integrated with pixel reconstruction, gradient consistency, and deep feature similarity. The performance of the proposed model is evaluated on the UCSD Ped2 and Avenue benchmark datasets. The proposed model achieves a significant improvement in term of frame-level Area Under Curve (AUC) score 92%. while maintaining computational efficiency suitable for real-time surveillance systems.